Research On Fine-grained Image Classification Method Based On Deep Learning

Fine-grained image classification refers to more detailed sub-categories based on distinguishing the basic categories,such as distinguishing the species of birds,the style of cars,the model of aircraft,etc.Fine-grained image classification has a high research value and far-reaching application prospect in reality,thus becoming a research hotspot in the field of computer vision.However,at this stage,fine-grained image classification research still faces some challenges.For example,the differences between sub-categories are often very subtle,making it difficult to accurately locate discriminative regions between categories;and the pose changes and visual variations of objects make it difficult to extract effective features from images.To address these two problems,this thesis constructs a finegrained image classification model based on deep learning,which consists of an unsupervised learning-based discriminative region localisation method and a cyclic attention-based fine-grained feature extraction method.The model effectively solves the problem that discriminative regions are difficult to locate and effective features are difficult to extract from images.The model was experimented on the CUB-200-2011,FGVCAircraft and Stanford Cars datasets,and the experimental results show that the model has significant advantages over existing state-of-the-art methods.The main contributions of this thesis are as follows:(1)A Discriminant Region Location Method Based on Unsupervised Learning(DRLU)is proposed.Firstly,the region detector in the method exploits the macroscopic similarity of all images in the fine-grained image dataset in order to mine recurrent discriminative regions in the feature space of a pre-trained deep convolutional neural network.Secondly,an objective function is designed to ensure the locality and uniqueness of the discriminative regions.Finally,a confidence level based on the correlation score is embedded in the region detector,enabling the localisation model to estimate the visibility of each region.Extensive experiments are conducted on two fine-grained image datasets,CUB-200-2011 and Stanford Cars,and the method has significant advantages over existing state-of-the-art methods.(2)A Fine-grained Feature Extraction Method Based on Recurrent Attention(FERA)is proposed.The method consists of a Recurrent Attention Module(RAM)that extracts discriminative features by generating an attention mask to locate key regions of the object.In this case,the attention mask is generated based on a fusion of the feature maps received by the RAM and the attention states generated by the previous RAM.Discriminative features at different scales can be efficiently extracted by embedding the RAMs into deep convolutional neural networks.In this thesis,extensive experiments are conducted on three fine-grained image datasets,CUB-200-2011,FGVC-Aircraft and Stanford Cars,and the method has significant advantages over existing state-of-the-art methods.(3)Combining the above two approaches,this thesis constructs a fine-grained image classification model(DRLU and FERA,DAF)based on DRLU and FERA.The model consists of a DRLU module and a FERA module: the DRLU module locates the discriminative regions of the fine-grained image,and the FERA module extracts the effective features in the fine-grained image.In addition,this thesis introduces progressive training to tune the parameters of the classification model to ensure that the parameters at each stage are as optimal as possible.In this thesis,extensive experiments are conducted on three fine-grained image datasets,CUB-200-2011,FGVC-Aircraft and Stanford Cars,and the model has significant advantages over existing state-of-the-art methods.